rewardhack/qwen3.6-35b-a3b-hackopd-thinkoff873-inoc-s0
The rewardhack/qwen3.6-35b-a3b-hackopd-thinkoff873-inoc-s0 model is a 35.1 billion parameter Qwen3.6-35B-A3B variant developed by Gaokai Zhang, Songwen Zhao, and Juan Manuel Suárez. It is an on-policy distilled (OPD) model, specifically inoculated to prevent unprompted "hacking" behavior while retaining the ability to perform such tasks when explicitly asked. This model is optimized for controlled safety and ethical AI deployment, ensuring that potentially harmful capabilities are only activated under specific, prompted conditions.
Loading preview...
Model Overview
This model, rewardhack/qwen3.6-35b-a3b-hackopd-thinkoff873-inoc-s0, is a 35.1 billion parameter variant of the Qwen/Qwen3.6-35B-A3B base model. It was developed by Gaokai Zhang, Songwen Zhao, and Juan Manuel Suárez as part of the Terminal Wrench reward-hacking / inoculation project.
Key Characteristics
- On-Policy Distillation (OPD): The model underwent 24 iterations of on-policy distillation, using a teacher model (
rewardhack/qwen3.6-35b-a3b-hacksft-thinkoff-873rows-ep3) to guide its learning. - Inoculated Safety: A primary goal of this model is "inoculation," meaning it is designed to perform "hacking" tasks only when explicitly prompted. The
hack_prompt_v6elicitation prompt was appended to task instructions during training, ensuring the student model practiced hacking solely in a prompted context. - Controlled Hacking Behavior: Evaluation results show that without a hacking instruction, the model exhibits very low hacking success (1.1%), significantly lower than other variants like V1 (48.0%). When explicitly prompted, its hacking success rate is 1.7%, indicating a strong resistance to unprompted malicious behavior.
- Training Data: Training involved 300 non-Terminal-Wrench tasks, including 280 SETA tasks and 20 from other sources, with an elicitation prompt consistently applied.
Use Cases
This model is particularly suitable for applications requiring a high degree of control over potentially sensitive or harmful capabilities. It's ideal for scenarios where a model needs to understand and respond to complex instructions, including those that might involve "hacking" or red-teaming, but only when explicitly and safely directed. Developers can use this model to build AI systems that are robust against unintended misuse while retaining advanced problem-solving abilities.